Benchmarking missing-values approaches for predictive models on health databases.

Area of Science:

• Machine Learning
• Data Science
• Health Informatics

Background:

• Large databases, common in health informatics, often contain missing values, complicating data management and analysis.
• Existing research on handling missing values primarily focuses on inferential statistics, not predictive modeling.
• Machine learning models, particularly discriminative approaches, offer new strategies for addressing missing data in large datasets.

Purpose of the Study:

• To systematically benchmark missing-value strategies for predictive modeling using large health databases.
• To compare the performance of native handling of missing values versus imputation methods in machine learning.
• To evaluate prediction accuracy and computational efficiency of different missing-value strategies.

Main Methods:

• Conducted a benchmark study on six large health datasets (electronic health records, brain imaging, surveys).
• Utilized gradient-boosted trees to compare native missing-value handling against simple and advanced imputation techniques.
• Assessed prediction accuracy and computational time for each strategy.

Main Results:

• Native handling of missing values within gradient-boosted trees demonstrated robust, fast, and accurate predictive performance.
• Imputation methods, while potentially improving prediction, incurred significantly longer computational times on large datasets.
• The inclusion of indicator columns for imputed values was crucial, suggesting data were not missing at random.

Conclusions:

• Supervised machine learning models with native support for missing values provide superior prediction accuracy with lower computational cost compared to imputation.
• When imputation is employed, adding indicator columns to denote imputed data is essential for optimal performance.
• Learning algorithms that incorporate missing values directly (missing incorporated attribute) are efficient and effective for large-scale health data.